AI can't yet 'find all vulnerabilities' with one prompt — expertise still decides bug-hunting output
moyix · x · 2026-10-22
dyn argues you still can't just prompt an AI to 'find all the vulnerabilities' — skilled researchers, custom harnesses, and other techniques are what's changing bug depth while driving down cost. moyix adds that the difference is hard to pin down: an expert casually chatting with a frontier model still outperforms a novice on the same model.
The takeaway: AI is lowering the cost of security research, but domain expertise and engineered tooling remain the decisive variables.
Related event: AI Vulnerability Research Still Depends on Expert Skill(3 posts)→
More from coding & agent
- Test LLM prompts like unit tests: Promptfoo turns evals into reviewable YAML — KhuyenTran16 · 2026-10-22
- Build a Grok agent that IS your ideal customer to pre-test content and offers — EXM7777 · 2026-10-22
- Point your coding agents at Clean Code instead of rule lists — it works — Al_Grigor · 2026-10-22
- RAG bot that refuses outside its docs scores 9/10 on held-out set — what broke — ZtasnimS75 · 2026-10-22
- REA: Open-Source Prompt System That Lets AI Agents Reverse-Engineer Any App or Website — blaizedsouza · 2026-10-22
- KIT lets AI coding agents formally verify code as they generate it — LingmingZhang · 2026-10-22